Join GitHub today
GitHub is home to over 28 million developers working together to host and review code, manage projects, and build software together.Sign up
Regression from 2.4.3 to 2.4.4 #185
I have the following piece of code:
I'm essentially taking the difference of two random vectors and finding that after repeating that for some time the difference changes significantly.
This code runs fine with numexpr-2.4.3, I'm using Anaconda on Windows AMD64 with the newest version of NumPy (1.10, 1.9 showed the same problem though), and the problem only appears when running multithreaded. Can someone else reproduce this?
It seems like a threading issue that leads to parts of the difference array being values very close to zero (~1e-315).
According to this https://www.diffchecker.com/vufifls8 (left git version of
I saw that the only non-trivial change between 2.4.3 and 2.4.4 is f6462f5, could this be the reason for this? It seems to rely on the value being set ...
It does indeed look harmless, doesn't it? The threads are synchronized after this so this is called exactly once instead of once per thread in the previous version. Maybe the synchronization is failing on Windows? A friend of mine tried the snippet on Gentoo and couldn't reproduce the error.
referenced this issue
Oct 29, 2015
Sorry, I missed the notification e-mails for this.